ANN FOR STRUCTURAL HEALTH MONITORING OF RC BUILDINGS: EXPERIMENTAL VALIDATION

Abstract

One of engineering requirements after a big earthquake is rehabilitation of damages in structures. Further, these damages can be visible or of invisible nature, and / or sometimes, it is difficult to reach the location for the manual inspection for damages in the structure. Therefore, there arises a need of intelligent means for the detection of these damages and finding out the extent of it. There are many available techniques for damage detection, which are under research evaluation, with some of them utilized in combinations to solve the issue which forms a part of Structural Health Monitoring (SHM). One technique of damage detection is through the use of Artificial Neural Networks (ANNs). In certain ANN based damage - detection technique, the modal parameters for different damaged conditions of the structure are given as inputs, and that of damaged conditions in the structure are given as outputs, for the training of the ANN models. Three algorithms for the detection of the damages have been proposed, with the concept of the correlation of the dynamic characteristics of the structure obtained from the vibration characteristics of it before and after the damage through its System Identification (SI), which involves the use of ANN. These algorithms are given, for three types of data available after or during the earthquake, based on Ambient Vibration Tests (AVTs), Free Vibration Tests (FVTs) and Shake Table Tests (STTs). The focus of study is towards the damage detection in Reinforced Concrete (RC) buildings, owing to the complexity of damages in these. More, results of proposed algorithms needed a validation. Whence, an Experimental model as a 1 : 4 scaled - model, of a hypothetical prototype of an earthquake - resistant RC building consisting of four - storey and 2 bays x 1 bay, has been designed and constructed in the laboratory. The direction along two bays and one bay of the model has been denoted as X direction that was the stiffer direction and the Y direction that was the flexible direction, respectively. This experimental model has been called as “Test model” in the thesis. It consisted of bracings in the exterior panels to induce different levels of damages with the special arrangement provided in it. The numerical model, called as a Finite Element (‘FE’) model in this thesis, of the same has been made. For the Validation data, tests were conducted on the Test model with different selected Test - damage combinations in it that led to set of vibration data from it. Further, the Operational Modal Analysis (OMA) of the Test data obtained from AVTs and FVTs has been made and Test - modal iv frequencies and Test - modal parameters have been found for different combinations of Test damages. Similarly, a MATLAB code has been used to find the Test - Frequency Response Functions (FRFs) for different combinations of Test damages from the Test data corresponding to STT. The models for conducting OMA have been called as “TEST model” in this thesis. Next, as per proposed algorithms, FE models has been updated corresponding to Test - modal parameters as obtained from AVT, FVT and STT data to yield the similar modal property as that obtained from the Test model for its non - damaged condition used in present study. The eigen - sensitivity based technique of updating of FE - model has been utilized for the mode - shape based updating of the FE model based on Test data from the AVTs and the FVTs. The mode shapes analyzed from the FE model has been compared with that of the TEST model to obtain an updated - FE model for each of the considered methods. The comparison of FRF has been made of the FE model with the FRF of the Test model obtained with the MATLAB code from the STT data. The modulus of elasticity ‘E’ values of the columns of the FE model have been selected as parameter to update in case of AVT and FVT based updating of the FE model. The update of ‘E’ value has been chosen, among the other parameters on which the stiffness depends, as the variation of ‘E’ values will not lead to change of other properties of the model, like, mass and geometrical properties. Also, the updating of stiffness of columns has been chosen, as updating of other component, like bracings, might have led to the failure of the damage detection algorithm. Further, the updating of only modal damping led to the updating of the FE model based on comparison of FE - FRF with the Test - FRF. Further, the updated - FE models have been obtained for AVT and FVT for Test data of X direction and Y direction. And, the updating of FE models with the data of STT has been made with respect to Y - direction only. Next, as per proposed algorithms, Latin Hypercube Sampling (LHS) technique has been used to obtain distributed samples of damages in terms of ‘E’ values obtained within the lowest bound value of 0 GPa to the highest bound value of 200 GPa, of the bracings chosen to have different combinations of damages, targeted to be found at the Floor (F) level and Panel (P) level, in the updated - FE models. In order to induce the damages in the different updated - FE models, the ‘E’ values of bracings were changed according to the set of 250 and 450 numbers of values from LHS damage sets for Y direction models and X direction models, respectively. Thereby, each of the updated - FE model yielded 250 and 450 numbers of models, for the respective Y direction updated - FE models and X - direction updated FE models, with different configurations of v damages in the bracings of the models. Each of these models with LHS sets of damages yielded the respective modal - parameters, i.e., mode frequencies and the corresponding mode shapes from the updated - FE models based on AVT and FVT, and that of FRFs from the updated models based on STT. Lastly, as per proposed algorithms, ANN models have been trained with inputs as the modal parameters and outputs as the LHS - damage sets obtained from the corresponding updated - FE models. These ANN models are expected to yield values of damages at different panels or floors of the Test model, when the set of modal parameter obtained from testing of structure after the event of such damages are given as input to them. The last part of this work was to validate the proposed algorithms. Therefore, trained models of ANN obtained were then given as the query to detect damages in the model corresponding to the modal parameter obtained from the Tests. The performances of ANN models corresponding to all the considered methods have been evaluated and the comparison has been made of their performances of detection of damage with respect to P level and F level, with respect to X and Y directions, and with respect to the AVT, FVT and STT method. First finding from the work is that the results from the comparison of the OMA of the Test model and a calibration model to calibrate OMA showed that the FDD peak picking for the simpler flexible - structure with lesser damping is easy as compared to that of the complex rigid - structure with more damping. Further, it has been perceived that the FDD peak picking technique in OMA is more reliable when the modes of vibration of structure are having lesser natural frequencies as compared to those modes of vibration that possess higher natural frequencies. Therefore, the number of mode shapes to be considered for damage detection depends on the type of structure. Second finding from the work is the observation that variations in modal frequencies differed relatively more in STTs. This may be attributed to the possibility of straining of the structure during STTs. Third finding from the work is that the STT method for stiffer direction failed at the stage of updating of the FE model based on FRFs due to mismatch of FE and Test FRFs for their substantial modal frequency variation. Fourth finding from the work is the accuracy of damage detection with STT based ANN model is obtained lesser as compared to ANN models based on AVT and FVT Cases. Damage predictions vi from ANN model along the more flexible direction in the plan of model are better as compared to that of stiffer direction. Floor level damages are predicted with better accuracy as compared to the prediction of panel level damages. Fifth, the most significant findings corresponding to the efficiency of different ANN models for detection of damages at F and P level are compiled in the last Chapter of the thesis. Further, it has been found that the proposed algorithms for the damage detection worked well. More, the comparative evaluation has shown varying performances of ANN models with satisfactory results for most of ANN models.